Seven Questions About General Agent Platforms

Jiang Wu · 2025-04-30

*Originally published in Chinese on April 30, 2025.*

This is my answer to seven questions about general agent platforms such as Manus, GenSpark, and Flowith, written in late April 2025.

What problem does the "general" part have to solve?

By the state of things now, everyone is still building platforms, looking for users, and setting up workflows. These entry points are all reasonable at the moment, but a general agent has to solve much more than this.

The first problem at this stage is that AI has to be useful. For AI to be useful, we need to find, in people's working life, the workflows AI can simplify, take from those workflows the user's needs in real scenarios, and from the needs find the places that can be automated.

Where does the opening for general agent platforms come from? I think it can be compared to the surge of the DeepSeek technology system. This wave truly completed the education of users' mindset from trying AI to using AI.

So, judging by today's results, a model's base capability no longer decides an application's barrier to some extent. What matters more is how, within this user base, we improve users' existing workflows and business flows further. One possibility is that you pick a product you like today, break its business flow apart, and integrate it on a platform, and that could be a new opportunity. If such opportunities accumulate and combine various business flows, could that become a general agent platform? There is room to enter at this point today.

From the viewpoint of how products progress, AI first has to be useful or fun. Fun needs no explanation, with a pile of time-killing products competing for it. Useful, for most people, is still a higher-order need, but this higher-order need comes with a great willingness and ability to pay. So the platform should first solve this higher-order need, using a combination of structures to make up for the limits and lack of intelligence of any single product.

This is the starting point at which the proposition holds today, and the initial motive for people to free up their time and spend more on AI. One regrettable but real fact is that, given users' current understanding, AI is not a pure necessity either as fun or as useful. Very likely the users who found it fun are the same users who found it useful, and AI has not yet made the essential jump from fun to useful. That is the opportunity for the next order of magnitude of product. On the timeline, which product appears first and finds the possibility first is something to look forward to.

What do users care about most?

What does the user base care about most right now? To some extent it has finished learning what AI can do. Think of it this way: smartphones became widespread overnight, but once you hold a smart device, how to use it and what to do with it still have to be defined by product managers and the people who build products.

Users don't really care how strong the model is, or what sophisticated technology is behind it. What they need today is the simplest, most natural way to turn their needs into structured capability and into output that saves their own time and effort. Users have finished the education about willingness, but they have not yet formed complete product workflow habits.

To B and to C are two entirely different logics and paths

To B and to C follow different logics and paths.

To B: accuracy. The logic of selling to businesses is to provide a 100% accurate solution for definite scenarios, because tolerance for error is very low. What counts in this dimension is the accuracy of input and output, and it doesn't even need high automation, but on definite results there is no tolerance for error. This may be the groundwork of the to-B service logic. The product is only a starting point, and higher added value comes from building precise output services for businesses.

To C: flexibility. In the to-C logic users are relatively tolerant of errors, but they have more scattered and uncertain needs. So in the to-C dimension, flexibility in customization matters: opening the customization mechanism to users and getting a 60 or 80 out of 100 result already meets their needs to some extent. In the key factors of both to B and to C, the product's own logic is not first. The core is to break down their expectations for the result, give them a matching solution, and earn added value from it.

In products right now, the early stage doesn't have to define B or C completely. You can walk a while first, build a general component or base well, go as far as it goes, see who the core users in the scenario are, and then split or invest in the core accordingly. For to B in particular, suppose we all supply matching component products in one dimension. Scenarios, core data requirements, and security requirements differ, so what can be split out also differs a great deal. In any dimension, you have to understand the user's scenario from the whole product's extensibility.

Who is the winner for this stage?

It is still very, very early to say. If we compare an agent's capability to the tools and applications for solving users' scenarios, we are probably not even at the stage when the browser had just been invented. From the competitive environment everyone faces, the workflow reshaping AI will bring has only just begun, or has not begun yet, and the change it brings is a structural and revolutionary solution. At this stage there is no winner. Leading or winning now only gets you some admission tickets, or the start of some possibilities, in a stage that is extremely early.

Startups have found their openings and have begun to define scenarios, but the giants are also looking for their own scenarios. They just won't do uncertain things too early. They only need to hold their ground, set up their core value points and barriers, and wait for the industry system to erupt in full.

What are the key factors in competition?

The key factors of competition come in the following dimensions.

The first factor is who can make the user's high-frequency, definite scenarios basically usable. That amounts to turning agent capability into a general capability. Whatever it's called, it must solve a general problem for users, with high accuracy. This is the core entry point for big companies and startups alike today: whoever finds the scenario, owns the scenario, and builds a solution on it. Big platforms build the infrastructure and let developers build solutions on it, and single products build the solutions directly. This first factor is whether you can get the user's scenario.

The second factor applies if you haven't got the user's scenario but have some new solution ideas, whether using the user's local hardware or the service capability you build in the cloud. These are places where you can be clever. The core value of an agent in the future must not be limited to the user's entry point and carrier, or to whether it runs locally. Whether it's a web page, a browser, or a mini program doesn't matter, and even if the input is only voice, text, or even expressions and actions. These are all possibilities, and they don't define the user's behavior and path. In every scenario where the user needs to see something or get a result, deliver a satisfying result with everything you have. That is the second place a competitive result can come from, and a chance for early startups to try. If it doesn't scale, no matter. If it does, they can build the matching technical or product barrier along the way.

The third factor is how your product and service reach users effectively and precisely. Must you build your own channel center or user entry point? Can you use existing entry points to deliver the service, plugging into big platforms in China and exposing an API overseas? These can all be key competitive factors, and the core is to understand what the user's need point and delivery point are.

At this moment the strategic choice for a product still matters a lot. Most of the time many people lack the patience, or need to build their so-called core barrier. But at any time, for any product, your only barrier is the user.

For the user groups served, what is the core point of payment in a scenario?

For the user scenarios served, the core points of payment may come in a few modules.

Often when we think about a product's value, we don't have to define whether to charge according to the usual billing cycle. We should weigh the system of service value we give users: whether it saves them time, adds to their ability, or raises their efficiency. Different points of value to pay for matter at the early entry stage. A good product naturally has several kinds of value. For this kind of product, saving time, raising output, raising efficiency, and increasing concurrency are all worth paying for. The list can't be exhausted, and it can only be judged product by product.

What are the risks ahead?

For those building agent products or platforms now, what are the definite risks ahead? Really there may be only one core risk: failing to build an effective data barrier of their own. Today, when the barriers of product capability, model capability, and even platform capability are all low, the only thing that keeps users is your data. If you can't define your data acquisition ability and data barrier from the user's own point of view, then more and more tools will help users migrate better. So defining the user's data system and data logic well may be the only problem, or the biggest one to face. If a product ends up without solving it well, the risk is large.

Home · All articles · Topics · 简体中文